Say how the site gets the decode mode of the older rows

[devYRPauli]

Problem

Amendment 8 of the design spec says that every result records its decode side in server.quirk_flags. It also says that an empty flag set means unverified. On main, the 12 llama.cpp rows and the 14 Ollama rows have empty server.quirk_flags and decode_mode: "unknown". Under that rule, a reader classifies 26 of the 32 rows as unverified.

registry/decode-modes-v1.json cites this section (lines 248 to 267) as a source of its three preset mappings. The site links that registry as the "cited preset mapping". The Codex review on #1 flagged the same claim in the README. #2 rewrote the README, but the spec kept the claim.

Fix

Rewrite the five lines of the DECIDED bullet in place. The bullet now says:

• The site shows the decode mode of each row.
• Schema v3 can record the mode in server.decode_mode.
• The site reads the mode of the older Ollama and llama.cpp rows from registry/decode-modes-v1.json.
• An unmapped preset stays unknown.

This matches the README and the M7 spec. The file keeps its line count, so the registry citation still points at lines 248 to 267.

Testing

• cargo run -q -p willitcall -- site --results results/ --out <dir> on main generates 32 rows. index.html shows "documented preset mapping" for the 12 llama.cpp rows and the 14 Ollama rows. It shows "recorded by run" for the 6 mlx-lm rows.
• cargo test --workspace: 155 tests pass.
• No code or test reads the spec. git grep finds its path only in registry/decode-modes-v1.json.